fix: fail loudly on missing NVIDIA libs and .ko, improve mknod logging

build-nvidia-module.sh:
- Replace silent glob cp for libnvidia-ml/libcuda with find + explicit error
  if library not found in extract dir (catches installer layout changes)
- Fix circular symlink bug: don't create .so.1 -> .so.1 if versioned file
  is already named .so.1
- Verify .ko count > 0 after build, fail loudly if none produced
- Show lib cache in final summary

bee-nvidia:
- mknod failures are now logged with ewarn instead of silently suppressed
- If nvidia not in /proc/devices (no GPU hardware), log clearly and exit clean

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

iso/builder/build-nvidia-module.sh

@@ -109,22 +109,38 @@ for ko in "$CACHE_DIR/modules/"*.ko; do
     strip --strip-debug "$ko" 2>/dev/null || true
 done
 
-cp "$EXTRACT_DIR/nvidia-smi"         "$CACHE_DIR/bin/"
+cp "$EXTRACT_DIR/nvidia-smi"           "$CACHE_DIR/bin/"
 cp "$EXTRACT_DIR/nvidia-bug-report.sh" "$CACHE_DIR/bin/" 2>/dev/null || true
-cp "$EXTRACT_DIR/libnvidia-ml.so."*  "$CACHE_DIR/lib/" 2>/dev/null || true
-# libcuda stub needed by nvidia-smi at runtime
-cp "$EXTRACT_DIR/libcuda.so."*       "$CACHE_DIR/lib/" 2>/dev/null || true
 
-# Create soname symlinks required by nvidia-smi on Alpine (musl/glibc via gcompat)
+# Copy userspace libraries — use find to handle any versioning scheme (libnvidia-ml.so.X.Y.Z or .so.1)
 for lib in libnvidia-ml libcuda; do
-    versioned=$(ls "$CACHE_DIR/lib/${lib}.so."[0-9]* 2>/dev/null | head -1)
+    found=$(find "$EXTRACT_DIR" -maxdepth 1 -name "${lib}.so.*" | head -1)
+    if [ -z "$found" ]; then
+        echo "ERROR: ${lib}.so.* not found in $EXTRACT_DIR"
+        ls "$EXTRACT_DIR/"*.so* 2>/dev/null | head -20 || true
+        exit 1
+    fi
+    cp "$found" "$CACHE_DIR/lib/"
+done
+
+# Verify .ko files were actually built
+ko_count=$(ls "$CACHE_DIR/modules/"*.ko 2>/dev/null | wc -l)
+[ "$ko_count" -gt 0 ] || { echo "ERROR: no .ko files built in $CACHE_DIR/modules/"; exit 1; }
+
+# Create soname symlinks required by nvidia-smi on Alpine (musl/glibc via gcompat + libc6-compat)
+for lib in libnvidia-ml libcuda; do
+    versioned=$(ls "$CACHE_DIR/lib/${lib}.so."* 2>/dev/null | grep -v '\.so\.1$' | head -1)
+    [ -n "$versioned" ] || versioned=$(ls "$CACHE_DIR/lib/${lib}.so."* 2>/dev/null | head -1)
     [ -n "$versioned" ] || continue
     base=$(basename "$versioned")
-    ln -sf "$base" "$CACHE_DIR/lib/${lib}.so.1" 2>/dev/null || true
+    # Only create .so.1 if versioned file is not already named .so.1
+    if [ "$base" != "${lib}.so.1" ]; then
+        ln -sf "$base" "$CACHE_DIR/lib/${lib}.so.1"
+    fi
     ln -sf "${lib}.so.1" "$CACHE_DIR/lib/${lib}.so" 2>/dev/null || true
 done
 
 echo "=== NVIDIA build complete ==="
 echo "cache: $CACHE_DIR"
-echo "modules: $(ls "$CACHE_DIR/modules/"*.ko | wc -l) .ko files"
+echo "modules: $ko_count .ko files"
-ls -lh "$CACHE_DIR/bin/"
+ls -lh "$CACHE_DIR/bin/" "$CACHE_DIR/lib/"

iso/overlay/etc/init.d/bee-nvidia

@@ -52,20 +52,27 @@ start() {
     # Without /dev/nvidiactl nvidia-smi returns NVML_ERROR_LIBRARY_NOT_FOUND (exit 12).
     nvidia_major=$(grep -m1 ' nvidiactl$' /proc/devices 2>/dev/null | awk '{print $1}')
     if [ -n "$nvidia_major" ]; then
-        mknod -m 666 /dev/nvidiactl c "$nvidia_major" 255 2>/dev/null || true
+        mknod -m 666 /dev/nvidiactl c "$nvidia_major" 255 2>/dev/null \
+            && einfo "created /dev/nvidiactl (major $nvidia_major)" \
+            || ewarn "/dev/nvidiactl already exists or mknod failed"
         for i in 0 1 2 3 4 5 6 7; do
             mknod -m 666 "/dev/nvidia$i" c "$nvidia_major" "$i" 2>/dev/null || true
         done
-        einfo "created /dev/nvidiactl and /dev/nvidia{0-7} (major $nvidia_major)"
+        einfo "created /dev/nvidia{0-7}"
     else
-        ewarn "/dev/nvidiactl: nvidia not in /proc/devices — no GPU hardware?"
+        ewarn "/dev/nvidiactl: nvidia not in /proc/devices — no GPU hardware present?"
+        eend 0
+        return 0
     fi
 
     uvm_major=$(grep -m1 ' nvidia-uvm$' /proc/devices 2>/dev/null | awk '{print $1}')
     if [ -n "$uvm_major" ]; then
-        mknod -m 666 /dev/nvidia-uvm       c "$uvm_major" 0 2>/dev/null || true
+        mknod -m 666 /dev/nvidia-uvm       c "$uvm_major" 0 2>/dev/null \
+            && einfo "created /dev/nvidia-uvm (major $uvm_major)" \
+            || ewarn "/dev/nvidia-uvm already exists or mknod failed"
         mknod -m 666 /dev/nvidia-uvm-tools c "$uvm_major" 1 2>/dev/null || true
-        einfo "created /dev/nvidia-uvm (major $uvm_major)"
+    else
+        ewarn "/dev/nvidia-uvm: nvidia-uvm not in /proc/devices"
     fi
 
     eend 0